The author spent five months attempting to build a corporate AI agent development environment using GoClaw as a foundation. The project faced significant challenges once the system reached ten users, including performance degradation, frequent crashes, and difficulties managing long-term tasks.
The developer identified that existing agent frameworks often lack robust human-in-the-loop (HITL) capabilities. Essential corporate requirements, such as handling task dependencies, managing permissions, and ensuring process continuity during system updates or agent restarts, were not natively supported.
A key finding was the necessity of an 'attention budget' for agents. Rather than providing the entire history to every sub-agent, the system must prioritize relevant information to prevent performance loss and context noise, ensuring that resources are managed across the entire delegation chain.
The author ultimately decided to abandon the GoClaw-based approach to build a custom system. This new architecture focuses on four pillars: execution via Temporal for long-running tasks, observability of event history, strict control policies, and a structured mechanism for learning from past experiences.
The new system distinguishes between raw event logs and actionable knowledge. Knowledge is treated as a projection of experience with a defined lifecycle, allowing for validation, updates, or invalidation, rather than simply dumping historical data into the model's context.
